Context ‘Depression, fatigue, pain’ (DFP) and ‘depression, insomnia, pain’ (DIP) symptom clusters (SCs) have been proposed in cancer. These symptoms are common and co-occur, that is, they constitute clusters of patients rather than symptoms.
Objectives The following research questions were addressed: (1) What is the frequency of co-occurrence of two symptom groups (DFP and DIP) in advanced cancer? (2) What is the degree of symptom item association within each symptom group? (3) Were either of these symptom trios associated with prognosis?
Methods We reanalysed a symptom data set of 1000 patients with advanced cancer. We identified the frequency of co-occurrence of two symptom groups: DFP and DIP, using both prevalence and severity data. The symptom associations were tested by χ2 and Spearman correlations. We also determined whether either of these symptom trios were associated with a major biological outcome, that is, survival by time-to-event analyses.
Results (1) Although DFP and DIP co-occured in about a quarter of the population, they were not SCs, but rather patient clusters. (2) Many persons had only one symptom from any symptom pair, and correlation coefficients were low for all symptom pairs. (3) Neither DFP nor DIP were associated with survival.
Conclusions Neither DFP nor DIP symptom item combinations constituted a specific cancer SC contrary to prior reports. DFP co-occurred in 27% and DIP in only 20%. Additionally, these symptom combinations were not associated with a biological outcome, that is, poor prognosis. Patient subgroups identified by shared symptom experiences alone do not identify SCs.
- symptom clusters
- Received 23 March 2015.
- Revision received 5 November 2015.
- Accepted 17 December 2015.
- Published by the BMJ Publishing Group Limited. For permission to use (where not already granted under a licence) please go to http://www.bmj.com/company/products-services/rights-and-licensing/
Statistics from Altmetric.com
People with advanced cancer typically experience multiple symptoms from the disease and treatments.1 These are chronic, distressing and often moderate to severe in intensity.2 Symptoms affect quality of life, treatment tolerance, functional status and disease progression. Symptom research has begun to focus on the cluster concept. One definition of a symptom cluster (SC) is three or more concurrent symptoms related to one another, and may (or may not) have the same aetiology.3 This definition should be perhaps revised to include the strength of association, that is, degree of correlation between symptoms. SCs should, at a minimum, occur with greater predictability than merely by chance.4 People with specific SCs may also have different clinical outcomes than those with individual symptoms.5 The clinical significance of SC research is the potential to better understand disease pathophysiology and also develop management strategies targeted at multiple symptoms.6 Depression, fatigue, insomnia and pain are prevalent and debilitating cancer symptoms.6–8 A National Cancer Institute State-of-the-Science Consensus Statement suggested that the three most prevalent and undertreated cancer-related symptoms were depression, fatigue and pain.9 It concluded there was insufficient evidence to support a SC of depression, fatigue and pain, but that more research was required.
Two general approaches to SC identification have emerged in the literature.5 They can be identified based on a priori assumptions about a relationship among selected symptoms such as nausea and vomiting, that is, a clinical approach or second by statistical analyses.10 ,11 In the former, a proposed clinical basis for SC identification includes which ones consistently form groups. Two groups ‘depression, fatigue, pain’ (DFP) and ‘depression, insomnia, pain’ (DIP) have been proposed based on empirical data.12–15
One criticism has been their use of predefined, that is, selected symptoms. Certain symptoms are assumed to cluster a priori, and data analysed accordingly. The strength of association has been determined by correlation tests, or multiple regression analysis. This identifies homogeneous patient subgroups based on symptom prevalence or severity. Significant correlations might also coexist with other symptoms, but which were not included in such clinically predefined analyses. Cluster components might vary based on the symptom items collected and/or analysed.
By discounting comprehensive symptom assessment before analysis, the clinical approach offers a less explicit understanding of SC. Since the items chosen directly determine the cluster's symptom content, exclusion of particular symptoms biases results. Comprehensive symptom assessment and measurement of symptom burden are important pre-requisites.
The symptoms in the two proposed clusters are common problems, that is, groups might simply constitute cohorts of patients (rather than clusters of symptoms). This clinical approach also differs from a statistical analysis because individuals can experience multiple SCs. Co-occurrence of common symptoms does not automatically mean that symptoms are intimately related, that is, that the presence of one predicts the likelihood of another, which would constitute a true SC.
The second approach16 is statistical analysis of symptom relationships without any prior assumptions about possible inter-relationships. These data-driven methods, for example, cluster or factor analysis, may reveal hidden data patterns.17 We previously identified seven SCs in advanced cancer by hierarchical cluster analysis among 25 symptoms with prevalence >15%.16 That analysis did not find any significant relationships between the individual symptoms of DFP or DIP. This observation challenged the clinical utility of these two specific proposed SCs. Given the lack of even moderate correlations between these symptoms when analysed as SCs, we reanalysed the original data but from the standpoint of patient rather than the time clusters. Since symptoms of depression, fatigue, insomnia and pain were preselected we did not use data-driven methods. The χ2 and Spearman correlations analysed both prevalence and severity data. We then examined the DFP and DIP patient groups. A patient cluster was defined as a group of people with specific concurrent symptoms (based on prevalence or severity). We address the following questions:
What was the frequency of co-occurrence (patient clusters) of individual items in two symptom groups: ‘DFP’ and ‘DIP’ in advanced cancer?
What was the degree of association (statistical clusters) within each symptom group?
Were these symptom trios associated with a biologically definable and important outcome, that is, prognosis?
The study was approved by the Cleveland Clinic Foundation Institutional Review Board. This was a post hoc analysis of previously reported data16 ,18 ,19 with a descriptive retrospective design. Data were originally collected prospectively for clinical purposes on 1000 consecutive cancer referrals to the Cleveland Clinic Palliative Medicine Program. Inpatient and outpatient referrals were included (all non-cancer diagnoses excluded). The majority of patients were not receiving any antitumour treatment. Some were receiving palliative radiation therapy for symptom control. Data were collected with an empirical symptom questionnaire of 38 symptoms. Interviewers were trained in symptom assessment. Symptoms were graded as present or absent, and (if present) as mild, moderate, or severe by the patient. Interviews were performed by an experienced specialist nurse or a clinical specialist physician, and lasted 15–20 min. Demographic data, cancer diagnosis, Eastern Cooperative Oncology Group (ECOG) performance score (PS) were collected from the questionnaire, and survival data (where available) from the medical record.
For clarity, all symptoms are listed in alphabetical order. For comparisons of prevalence with severity, symptoms were all analysed first by their prevalence and then those rated as moderate/severe. Moderate or severe symptoms may be considered empirically to represent those most likely to be clinically important, and are referred to as such throughout. For DFP there was complete prevalence data (N=962) on all three symptoms, and 769 had complete severity data available. Survival data were available for 869. For DIP, there was prevalence data on all 3 in 964. Similarly, 806 had severity data on all three symptoms, and 725 of these also had survival data. These two data sets serve as the basis for this report.
The associations between the individual symptoms in ‘DFP’ and separately in ‘DIP’ were tested by χ2 test. Associations between the severity of symptoms were tested with Spearman correlations. Cox proportional hazards analysis assessed the univariable relationship of the DFP and DIP groups to survival postreferral. Kaplan-Meier survival curves were constructed for all four (pain, depression, fatigue, insomnia) symptoms (figures 1⇓⇓–4). Results were summarised as the hazard ratio (HR), 95% confidence interval (CI), and corresponding p value. HR>1 indicates a greater risk of mortality, while HR<1 a lower risk. Cox results were included on each figure. All statistical tests were two sided; p<0.05 indicated statistical significance. Analyses were performed with SAS software (SAS Institute Inc, Cary, North Carolina, USA).
Descriptive summary of the data
The demographic and clinical characteristics of the population have been reported in detail elsewhere.20 Table 1 contains descriptive information of the study population. From the database, prevalence data on all three symptoms were available for 962 patients for DFP and 964 patients for DIP. This data set has been the basis of several reports21–23 including some on SCs.16 ,18 ,19 Those with 0, 1, 2 or 3 symptoms are in table 2 (DFP), and in table 3 (DIP), along with the specific symptom combinations present. The median age was 65 years (range 12–94); 55% were male; 54% had ECOG PS 3–4. Most were Caucasian. Lung (24%), breast 9%, colorectal 8%, prostate 7% and unknown primary (6%) were the commonest cancer primary sites. The distribution of cancer diagnoses was consistent with estimated causes of cancer deaths in the USA in 2013.24
Depression, fatigue and pain
Table 2 presents symptom prevalence and severity data for DFP. Overall, complete symptom prevalence data was available for 962 patients, and in 921 at least one of these symptoms was experienced; 25% (N=242) experienced one of these symptoms (most often pain), 44% (N=418) two, and 27% (N=261) all three. Symptom severity was recorded for 769 patients; 642 had clinically important symptoms; 40% one, 34% two and 9% for all three symptoms, while 17% did not experience any moderate/severe symptoms. For prevalence data (table 4), χ2 was significant for pain-depression (p=0.04) and fatigue-depression (p<0.001). When we examined severity, the Spearman correlations for pain-depression were very low r=0.05 (p=0.13); pain-fatigue r=0.001 (p=0.97); depression-fatigue r=0.17 (p<0.001). Even though the depression-fatigue correlation was statistically significant, the value was again low (all r<0.2). For clinically important symptoms (n=769), the χ2 statistic was only significant for moderate/severe fatigue-depression (p<0.001). Severity associations among DFP are presented in table 5.
Depression, insomnia and pain
Table 3 presents symptom prevalence and severity data for DIP. Overall, complete symptom prevalence data was available for 964 patients, and in 897 at least one of these symptoms was experienced; 33% (N=318) experienced one of the three symptoms (most often pain); 40% (N=387) two, and 20% (N=192) all three. Symptom severity was recorded for 806 patients. When we examined severity, 630 had clinically important symptoms; only 7% had all three rated as clinically important; 26% had two, and 46% one, while 22% did not experience any moderate/severe symptoms. The Spearman correlations for prevalence (N=964) again were low r=0.08 (p=0.01) for pain-insomnia; r=0.06 (p=0.05) for pain-depression, and r=0.11 (p=0.001) for insomnia-depression (table 6). Spearman correlations for clinically important DIP (N=806) were r=0.07 (p=0.04) for pain-insomnia, r=0.05 (p=0.19) for pain-depression and r=0.13 (p<0.001) for insomnia-depression. Most associations were statistically significant. Many persons had only one symptom from each symptom pair. Like DFP, correlation coefficients were again very low for all symptom pairs (all r<0.2). When we examined clinical severity associations, the pain-depression correlation was not significant (r=0.04; p=0.26) (table 7). Correlations between insomnia-pain (r=0.13, p<0.001) and insomnia-depression (r=0.15, p<0.001) were statistically significant but low (all r values <0.2).
Neither the DFP nor DIP symptom groups were associated with worse survival. Pain was the only symptom associated with survival with actually a lower risk of mortality, that is, longer survival than in those without. This was seen for overall pain prevalence (figure 1), for clinically important prevalence, and also overall severity.
Specific SCs have been identified16 which occur with a very high degree of mathematical predictability. In our analysis, this did not apply to either the DFP or DIP symptom groups. There were three important findings from a new analysis:
Although ‘DFP’ and ‘DIP’ co-occured in about a quarter of patients with advanced cancer, they were not SCs, but rather patient clusters.
Many persons had only one symptom from any symptom pair, and correlation coefficients were low for all symptom pairs.
Neither DFP nor DIP was associated with survival in advanced cancer.
Clinical experience has suggested that some symptoms often co-occur; for example, the proposed ‘DFP’, ‘pain, confusion, constipation’, and ‘fatigue, insomnia, pain’ clusters.3 ,13 ,25 The disadvantage of this approach is the absence of proof of mathematical relationships between symptoms (that are highly prevalent anyway) and the limited numbers of symptoms included and examined. Moreover, significant correlations may also exist between other symptoms which were not included in analysis. Symptoms within a SC should have a stronger association with each other than items in other clusters. However, the criteria for such correlations have not been defined. Symptom relationships have been identified by correlation tests and multivariable analysis, for example, regression. Others have measured relationship based on the effect of symptoms on outcomes.3 Most agree that SCs have a greater adverse impact on outcomes than individual symptoms.18 ,26 In one study (for the preselected symptoms of pain, fatigue and insomnia), regression analysis suggested outcome variance, for example, functional status and correlations between symptoms were tested.3 Fatigue and pain (but not sleep insufficiency) were statistically significant predictors of functional status.3 Another study of 51 lung cancer survivors examined the impact of the proposed DFP cluster on quality of life by regression analysis.15 Fatigue was correlated with depression and pain. Both depression and fatigue had some relationship to quality of life. Pain, however, was not correlated with depression or quality of life. Common symptoms naturally often co-occur, but may not predict clinical outcomes.
Our data provides limited support for a (weak) link between pain-depression, but not pain-fatigue. In our data set, only 27% had all three symptoms, and in just 9% were they all clinically important. Similar to our data, one other study identified that fatigue and depression (but not pain) had some impact on quality of life in patients with ovarian cancer.12 Individuals with pain may be more prone to depression, because of its adverse effects on mood and physical function. Although fatigue and depression had the strongest associations (both for prevalence and severity), many people still had only one of the two symptoms. A possible DFP constellation might suggest a fatigue syndrome (associated with pain and depression), or a pain syndrome (that causes fatigue and depression).27 Individual symptom items can exist in isolation or in combination with others. Therefore, the presence of one does not predict the presence of the others in the ‘cluster’. Thus, in our analysis, this DFP symptom trio did not meet reasonable criteria for a SC.
DIP is also sometimes considered a SC.28 Our data again did not support this. We found that they co-occured in 20%; only 7% had all three symptoms (when rated as clinically important). When we examined both prevalence and severity associations for a possible DIP cluster, most (while statistically significant) again had low correlation coefficients for all symptom pairs (all r<0.2). This severely limits the validity of any proposed DIP cluster. Similarly, sleep disturbance (especially early morning awakening) is a symptom of depression. Poorly controlled pain may cause sleep problems and, in turn, daytime fatigue. Sleep abnormalities have been correlated with cancer fatigue. Sleep disturbance partially mediates the pain–fatigue relationship. This co-occurrence may be important for screening and treatment (especially for depression and fatigue), which are often under-recognised and undertreated. Patient (rather than symptom) clusters may help identify high-risk populations who need early intervention. However, this approach needs theoretical justification for selection of only the most common symptoms in cluster analysis. There is, by our analysis, insufficient evidence to support a specific DIP cluster.
We did not find any relationship between either the proposed DFP or DIP clusters and survival, that is, no biological effect, nor did these proposed SCs have a synergistic effect in that respect. Pain was the only symptom associated with longer survival, perhaps due to more incident pain from greater physical activity in those with better performance status. One study26 did suggest that a proposed DFP cluster was associated with disease progression in hepatobiliary cancer. They found that approximately 25% of patients experienced high levels of depression, fatigue and pain. Overall, the three symptoms were reported in 62–85% of patients from diagnosis to 6-month follow-up. The main limitation of that study was again that they selected only the most common symptoms in cluster identification. Moreover, patients (rather than symptoms) were divided into clusters. The study has not been replicated in other primary tumour sites. Since multiple cancer symptoms occur together, bivariate correlations alone are unlikely to describe their complex inter-relationships. Our data suggests that a purely clinical approach is ineffective in detecting subtle complex relationships among symptoms.5
Neuroendocrine-immune models might account for the frequent co-occurrence of depression, fatigue and pain in cancer.8 ,29 ,30 One study31 explored the relationship between a proposed DFP cluster and biological markers in advanced breast cancer. They concluded that elevated stress hormones predicted a DFP cluster. Although most associations were statistically significant, the small study sample (n=104) and the nature of the sample (only females with advanced cancer) are noteworthy. Moreover, the study tested a hypothesis with data collected for an observational study, and may not be ideal. Another study13 in breast cancer found significant associations between depression, fatigue and pain, and outcomes like total health status. This convenience sample was composed of a selected group of patients with breast cancer, and this limits any conclusions.
A proposed DFP cluster may reflect a sickness behaviour syndrome (pain, fatigue, cachexia, fever and anorexia) from antitumour treatments.32 This might be generated by the cancer and host-induced hormones and immunmodulators. However, this model cannot explain many other symptoms not included in sickness behaviours.33 Systemic inflammation may be another common underlying mechanism for a possible DFP cluster, perhaps mediated by proinflammatory cytokines (interleukin (IL) 1, tumour necrosis factor (TNF) α and interferon (IFN) α). This hypothesis was tested in one study.14 DFP clustered approximately two to four times more in individuals with cancer cachexia who experienced all three symptoms. DFP were chosen as they were common cancer symptoms. The presence of other symptoms and their potential role in such a cluster were not examined. Another study showed that certain proposed SCs were prevalent among community-dwelling adults with a cancer history, with DFP most prevalent.34 It found a positive association between SC and coexisting medical conditions. The influence of other factors (for example, cancer type, disease severity, treatments) that may modify the risk for DFP was not assessed. It is possible that DFP or DIP clusters may indeed exist in other non-cancer patient populations, for example, chronic benign pain, and this should be investigated. Significant associations do exist between other symptoms than DFP and DIP which are SCs and have clinical significance. For example, the fatigue/anorexia–cachexia, aerodigestive cluster, and debility clusters have clinically and statistically important negative impact on survival in advanced cancer.18 There would be other possible statistical approaches if we were trying to identify SCs, but the desire here was to assume that each set of three symptoms analysed were a cluster, and see how often they co-occur. If the symptoms represented really do cluster, we would expect most patients to either have none of the three or all three symptoms, and find strong associations between each pair of symptoms in the cluster, but we did not.
One of the strengths of this analysis of patient clusters is that multiple cancer primary sites were included, whereas some previous work6 focused on a single tumour site. Data was collected consecutively; patients were not selected, and the distribution of diagnoses closely resembled common causes of cancer mortality in the USA. There were several potential limitations to our analysis. The database is old but it is probably reasonable to presume little major change in cancer symptom prevalence over time (except perhaps for some persistent or novel treatment-related symptomatology). Symptom assessment was based upon systematic history-taking not a validated instrument. Inter-rater reliability was not assessed. We used one-item questions for symptom prevalence, and categorical scores to measure symptom severity. Studies have found that individual symptom ratings (like pain) have shown minimal difference compared to composite measures.35 ,36 DFP and DIP were examined as separate item groups. The data was cross-sectional. We did not collect data on disease stage, prior antitumour treatments, or physical illnesses at baseline, which might have biased certain outcomes of interest including survival. It is possible that clusters present earlier in the disease trajectory might alter with time or disease progression. Hence, our findings apply only to the setting of advanced cancer.
We examined the frequency of co-occurrence of two symptom groups: ‘DFP’ and ‘DIP’ in advanced cancer. The population was representative of USA cancer mortality. Neither DFP nor DIP combinations constituted a specific cancer SC contrary to prior reports. They sometimes co-occurred (people clusters), but that is different from a SC. DFP co-occurred in 27% and DIP in only 20%. These symptom combinations were not associated with worse survival unlike other known SCs. Construction of patient subgroups by similar symptom experiences alone does not identify SCs. Further research is needed to define SCs and the best methods to identify them.
- Received 23 March 2015.
- Revision received 5 November 2015.
- Accepted 17 December 2015.
Presented at the 3rd International Seminar of the Palliative Research Center and European Association for Palliative Care Research Network, Milan, Italy, 5–6 December 2013.
Harry R. Horvitz Center for Palliative Medicine is a World Health Organization Demonstration Project in Palliative Medicine and an ESMO Designated Center of Integrated Oncology and Palliative Care.
Contributors AA, DW and KH planned the study. LR conducted statistical analysis.
Competing interests None declared.
Ethics approval Cleveland Clinic IRB.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement Data are available for sharing on request.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.